Non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology

Abstract

The present invention discloses a dynamic monitoring and early warning system for safety risks in non-coal mines based on point-surface fusion technology, which relates to the technical field of mine safety monitoring. The system uses the four-dimensional spatiotemporal coordinate data of surface monitoring points to construct a dynamic surface model; adopts a deep learning large model (DSIM) to dynamically invert a probabilistic three-dimensional underground geomechanical model from the surface and quantify the uncertainty; adaptively evolves the DSIM model online by comparing the difference between predictions and actual observations; integrates a causal inference engine to deeply trace the risk driving factors; and finally generates an intelligent early warning with clear causal attribution. The present invention realizes underground state perspective, risk cause tracing and model self-evolution, significantly improving the accuracy, reliability and intelligent decision-making support level of monitoring and early warning.

Description

Technical Field

[0001] The present invention relates to the technical field of mine safety monitoring, and in particular to a non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology. Background Art

[0002] As mine safety production becomes more and more important, although the existing monitoring and early warning technologies use a variety of sensors, they often have problems such as independent data operations, low degree of integration, insufficient deep information mining, rigid early warning models, and difficulty in adapting to dynamic changes in mines. This leads to low efficiency in the use of monitoring data, making it difficult to accurately and timely identify and warn of major safety risks, especially in the prediction of risk evolution under complex geological conditions and mining disturbances. Effective dynamic monitoring and precise early warning are crucial to preventing and resolving major safety risks.

[0003] In order to improve the early warning capability of safety risks in non-coal mines, some explorations have been made in existing technologies, such as:

[0004] Chinese invention patent CN116805213B discloses an early warning method and system for comprehensively assessing safety risks. The method obtains historical data from non-coal mining enterprises, classifies and cleans it, establishes static and dynamic safety risk models, inputs periodic data to generate risk results and determine risk levels, and finally generates an assessment report and issues early warnings. The patent mainly focuses on using historical data for periodic risk assessment and grading.

[0005] Chinese invention patent CN117057601B discloses a non-coal mine safety monitoring and early warning system based on the Internet of Things. The system collects multi-source data through the Internet of Things to a big data platform, establishes a risk indicator system and a safety accident evaluation model, and inputs real-time indicator values ​​into the model to obtain early warning risk values, aiming to provide timely early warnings. This patent focuses on building an indicator system and risk assessment based on current indicator values.

[0006] The above designs have made some progress in non-coal mine safety risk assessment and early warning by analyzing historical data to establish risk models (such as CN116805213B) or using the Internet of Things to build an indicator evaluation system (such as CN117057601B). However, there are still certain limitations, such as: lack of in-depth understanding of the physical state inside the mine, shallow levels of data fusion and information mining, insufficient model adaptability, and limited risk tracing and decision support capabilities.

[0007] Therefore, there is an urgent need for a new dynamic monitoring and early warning system for non-coal mine safety risks that can deeply integrate point and surface monitoring data, intelligently invert underground conditions, have adaptive evolution capabilities, and trace risk causality, in order to overcome the shortcomings of existing technologies. Summary of the Invention

[0008] The purpose of the present invention is to overcome the deficiencies of the prior art and propose a non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology to solve the above-mentioned problems.

[0009] The object of the present invention is achieved through the following technical solutions: a non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology, comprising a data processing module, a dynamic surface modeling module, an underground state inversion module, a prediction and difference analysis module, a model self-evolution module, a causal inference engine module, and a risk assessment and early warning module;

[0010] The data processing module includes at least: a data acquisition unit for collecting non-coal mine monitoring data; a data preprocessing unit for preprocessing the monitoring data to obtain a standardized four-dimensional spatiotemporal data set; a dynamic surface modeling module connected to the data preprocessing unit of the data processing module for constructing a dynamic surface geometric model sequence representing the temporal changes of the surface morphology based on the four-dimensional spatiotemporal data set;

[0011] The subsurface state inversion module is connected to the dynamic surface modeling module. It is used to perform inversion inference from surface dynamics to subsurface state based on a sequence of dynamic surface geometric models using a deep learning large model to obtain a probabilistic three-dimensional subsurface geomechanical model at the current time step. This model includes estimates of subsurface parameters, stresses, structures, and related uncertainties.

[0012] The prediction and difference analysis module is connected to the underground state inversion module and the dynamic surface modeling module; the prediction and difference analysis module at least includes: a forward prediction unit for predicting a predicted surface geometry model at the end of a preset time interval in the future based on the currently inverted underground model; a difference calculation unit for obtaining the actual observed surface geometry model at the end of the preset time interval in the future and calculating the prediction-observation difference between the predicted surface geometry model and the actual observed surface geometry model;

[0013] The causal inference engine module connects the dynamic surface modeling module, the subsurface state inversion module, the prediction and differential analysis module, and the module for acquiring environmental / activity data;

[0014] The risk assessment and early warning module connects the underground state inversion module, the prediction and difference analysis module and the causal inference engine module.

[0015] The monitoring data includes at least four-dimensional spatiotemporal coordinate data of surface monitoring points containing three-dimensional spatial coordinates and corresponding timestamps; and optionally collects surface morphology data, environmental factor data and engineering activity data; preprocessing includes cleaning, denoising, coordinate unification and time synchronization processing. The deep learning large model built into the underground state inversion module is a spatiotemporal graph variational autoencoder architecture that integrates physical information. The architecture includes a spatiotemporal graph encoder for processing the time series of dynamic surface geometric models, and a conditional variational autoencoder for generating a probabilistic three-dimensional underground geomechanical model.

[0016] The causal inference engine module includes at least: a causal graph maintenance unit for maintaining or dynamically learning the graph structure representing the causal relationship between variables; a causal effect quantification unit for quantifying the degree of causal influence of risk-related variables; and a risk attribution analysis unit for identifying key causal drivers that lead to changes in risk status or abnormal phenomena.

[0017] The probabilistic three-dimensional underground geomechanical model output by the underground state inversion module includes at least: mean estimation and uncertainty quantification provided for the underground medium mechanical parameter field, three-dimensional stress field, and three-dimensional strain field, and the quantification is at least expressed through the variance field; and the probabilistic position representation provided for the geological structure.

[0018] The conditional variational autoencoder that integrates physical information embeds physical law constraints related to geomechanics through loss functions or model structures. The constraints are used to punish inference results that do not conform to physical laws during training or inference. Physical laws at least include mechanical equilibrium equations or constitutive relations; the risk assessment and early warning module at least includes: a comprehensive assessment unit for integrating the inverted underground state and its uncertainty output by the underground state inversion module, the predicted future risk and model prediction reliability indication output by the prediction and difference analysis module, and the key causal driving factors identified by the causal inference engine module to perform a comprehensive safety risk level assessment; the early warning generation unit is used to generate an early warning containing risk level and causal attribution information based on the comprehensive safety risk level assessment.

[0019] The dynamic surface modeling module uses a dynamic triangulation algorithm or adaptive grid generation technology to construct a dynamic surface geometric model based on three-dimensional spatial coordinates and four-dimensional spatiotemporal coordinate data with corresponding timestamps; the model self-evolution module connects the difference calculation unit of the prediction and difference analysis module and the underground state inversion module to update the model parameters of the underground state inversion module online based on the prediction-observation difference.

[0020] The model self-evolution module uses the prediction-observation difference as an error signal to adjust the model parameters of the subsurface state inversion module through an online learning algorithm, and the algorithm includes at least one of gradient-based updating or reinforcement learning.

[0021] The causal inference engine module is further used to: learn or update the causal relationship graph in its causal graph maintenance unit using a constraint-based or scoring-based causal discovery algorithm; and quantify the causal effect through its causal effect quantification unit using an intervention calculus or counterfactual model-based method, and perform attribution analysis through its risk attribution analysis unit.

[0022] The comprehensive assessment unit of the risk assessment and early warning module uses preset rules, fuzzy logic or meta-learning models to perform weighted fusion or logical judgment on the state and uncertainty information from the underground state inversion module, the predicted risk and model deviation information from the prediction and difference analysis module, and the key causal driving factor information from the causal inference engine module to determine the final comprehensive risk level.

[0023] The warning information generated by the warning generation unit of the risk assessment and warning module, in addition to the risk level, location and time elements, also clearly includes a causal attribution text description of the most important one or more causes leading to the current risk status.

[0024] The data acquisition unit of the data processing module is also used to collect surface state monitoring data obtained through at least one of ground-based InSAR, UAV LiDAR or UAV photogrammetry technology, and the dynamic surface modeling module and / or underground state inversion module is further used to integrate or utilize surface monitoring data as supplementary information or constraints when constructing a surface model or inverting the underground state.

[0025] The beneficial effects of the present invention are:

[0026] 1. Unlike traditional methods that rely solely on surface displacement or shallow-layer information, the present invention utilizes a subsurface state inversion module. Based on dynamic surface monitoring data (core consisting of four-dimensional spatiotemporal data, with optional fusion of surface data), it can intelligently infer three-dimensional subsurface geological structures, geotechnical parameter fields, and stress / strain state distributions that are difficult to directly detect. This allows for a deeper understanding of the inherent mechanisms of deformation and the sources of risk. The inverted subsurface model is probabilistic, providing not only an optimal estimate of the subsurface state but also quantifying the degree of uncertainty in this estimate. This enables decision makers to make judgments based on risk and confidence rather than relying on a single deterministic result. By embedding physical law constraints, the inverted subsurface model is ensured to conform to basic geotechnical mechanics principles, avoiding the potential inferences that violate physical common sense that may be produced by purely data-driven models, thereby improving the model's reliability and engineering application value.

[0027] 2. The system not only assesses the current state but also makes forward-looking predictions based on the current inverted underground model, estimating deformation trends or risk evolution over the next period of time, providing a time window for the formulation of preventive measures. Through the "prediction-observation-comparison-model correction" mechanism, the core underground state inversion model can continuously learn from actual monitoring data, adjust its own parameters online, and automatically adapt to the dynamic evolution of geomechanical behavior caused by factors such as mining and environmental changes in the mine. It can maintain long-term monitoring and early warning accuracy and robustness, far superior to traditional models with fixed parameters.

[0028] 3. The introduction of a causal inference engine module can use algorithms (such as constraint / scoring-based causal discovery, intervention calculus, and counterfactual models) to analyze the complex relationships between multi-source data, distinguish between correlation and causality, and identify the fundamental driving factors that lead to changes in risk status or abnormal phenomena. For example, it can distinguish whether it is caused by heavy rainfall, the impact of excavation activities, or the deterioration of internal structures. It can not only qualitatively determine the cause, but also quantitatively estimate the contribution of different factors to the risk, providing an unprecedented scientific basis for understanding the primary and secondary contradictions of risks and formulating precise intervention measures.

[0029] 4. Risk assessment no longer relies solely on the threshold of a single indicator. Instead, it intelligently integrates multiple dimensions of information, including underground conditions (including uncertainty), future predicted risks, the reliability of the model itself (reflected by prediction-observation discrepancies), and key causal drivers. The assessment results are more comprehensive, smarter, and closer to the essence of the risk. Early warning information clearly includes causal attribution explanations, directly informing managers of the "cause of the risk", greatly improving the comprehensibility and decision-making guidance value of early warning information. Managers can quickly determine what targeted measures should be taken (such as strengthening drainage, adjusting operations, and strengthening engineering), making emergency responses more timely and effective.

[0030] 5. The system not only processes core four-dimensional spatiotemporal point data, but also effectively integrates surface monitoring data such as ground-based InSAR and UAV LiDAR, achieving a broader point-to-surface integration, improving the integrity and accuracy of surface deformation field characterization, and thereby enhancing the accuracy of underground inversion and risk assessment. It integrates data acquisition, processing, modeling, inversion, prediction, self-evolution, causal analysis, assessment and early warning within a unified system framework, achieving process automation and intelligence. Compared with traditional decentralized and highly manual monitoring and analysis models, the efficiency and overall effectiveness of monitoring and early warning are greatly improved. The entire system has been upgraded from a traditional monitoring tool to a comprehensive mine safety risk management platform with deep insight, cause tracing, forward-looking prediction and intelligent decision support capabilities, providing strong technical support for non-coal mines to achieve more scientific, proactive and effective production safety management. BRIEF DESCRIPTION OF THE DRAWINGS

[0031] Figure 1 The system architecture of the present invention Figure 1 ;

[0032] Figure 2 It is a timing diagram of the present invention;

[0033] Figure 3 The system architecture of the present invention Figure 2 . DETAILED DESCRIPTION

[0034] The technical solutions of the present invention will be clearly and completely described below in conjunction with the embodiments. Obviously, the described embodiments are only part of the embodiments of the present invention, rather than all the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative work are within the scope of protection of the present invention.

[0035] Example 1:

[0036] like Figures 1 to 3 As shown, the core system architecture and basic monitoring and early warning process of this embodiment provide a basic implementation method of a dynamic monitoring and early warning system for safety risks in non-coal mines based on point-surface fusion technology, aiming to establish a core framework that can dynamically monitor, preliminarily invert underground conditions and conduct risk early warning.

[0037] The system of this embodiment is deployed in a typical non-coal mining environment, such as the slope or spoil dump area of ​​an open-pit mine. The system hardware deployment is as follows: several high-precision GNSS receivers are deployed at key locations on the surface of the mining area (such as the top and foot of the slope, and areas with obvious deformation) as the main sensors of the data acquisition unit, which are used to continuously obtain the three-dimensional spatial coordinates and time information of the monitoring points. Optionally, environmental sensors such as rain gauges are deployed in or near the site, and their data is also collected by the data acquisition unit. One or more servers are set up as the system operation platform, and subsequent data processing, modeling, analysis and early warning software modules are deployed.

[0038] The data acquisition unit of the data processing module is responsible for receiving the raw observation data transmitted by each GNSS receiver and other optional sensors in real time through the network interface.

[0039] The data preprocessing unit performs the following operations on the received raw data:

[0040] Data cleaning removes obvious gross errors and abnormal signals; coordinate conversion converts geodetic coordinates into unified mine engineering coordinates; time synchronization aligns data from different sensors at preset time intervals (for example, 30 minutes) and outputs a standardized four-dimensional spatiotemporal coordinate dataset containing (X, Y, Z, T).

[0041] The dynamic surface modeling module receives the standardized four-dimensional spatiotemporal dataset output by the data preprocessing unit. In this embodiment, the module uses a dynamic triangulation algorithm (such as a real-time updated Delaunay triangulation) to construct and update the surface triangular mesh model based on the coordinates of the valid monitoring points at each time step, and outputs a dynamic surface geometric model sequence representing the real-time morphology and deformation of the surface.

[0042] The dynamic surface modeling module uses the following surface dynamic activity index formula:

[0043]

[0044] illustrate:

[0045] : surface dynamic activity index;

[0046] : current time step;

[0047] : Surface mesh vertex i at time step T k Velocity vector inside;

[0048] : Indicates averaging all vertices;

[0049] : represents the magnitude (or size) of the vector;

[0050] : Indicates the spatial standard deviation of velocity;

[0051] and β: weight coefficient, used to balance the influence of the two parts.

[0052] The subsurface state inversion module (DSIM) receives the surface model time series output by the dynamic surface modeling module. In this embodiment, DSIM uses a pre-trained deep learning large model (which can be based on the infrastructure of CNN, RNN or their combination). The core function of this model is to invert and infer the underground geomechanical state based on the input surface deformation sequence, and its output is a probabilistic three-dimensional underground geomechanical model.

[0053] On a preset 3D grid, it provides the mean field (as the best estimate) and variance field (as the uncertainty measure) of key mechanical parameters (such as elastic modulus, cohesion) and stress / strain states of the underground medium, as well as the probabilistic location representation of major geological structures (such as potential slip surfaces).

[0054] The physical constraints embedded in the Subsurface State Inversion Module (DSIM) are formulated using the following uncertainty-weighted physical loss function:

[0055]

[0056] illustrate:

[0057] : uncertainty-weighted physical loss function;

[0058] : represents the average of all sampling points in the three-dimensional domain;

[0059] :The physical governing equations at point the residuals of (calculated by automatic differentiation);

[0060] : DSIM output fields (such as stress fields ) the predicted variance at that point (from PI-CVAE);

[0061] : A small constant that prevents the denominator from being zero.

[0062] The data fusion application of the Subsurface State Inversion Module (DSIM) uses the following multimodal data fusion constraint formula:

[0063]

[0064] illustrate:

[0065] L MMDFC : Multimodal data fusion constraints;

[0066] :By generating the model Three-dimensional surface deformation field obtained by FFS simulation

[0067] :Will Projection to InSAR line of sight

[0068] : Observed InSAR deformation field

[0069] : In common coverage area The L2 norm squared calculated on

[0070] : Sample the simulated points corresponding to the UAV point cloud positions from the simulated deformation field

[0071] : Observed drone point cloud or its differential data

[0072] : In the area Distance metrics between point sets on (such as root mean square error)

[0073] and : Respective weight coefficients

[0074] The forward prediction unit of the prediction and difference analysis module uses the underground model mean field output by the underground state inversion module as input, and predicts the surface deformation at the end of a preset time interval in the future (such as the next monitoring cycle) through a simplified and fast-computing forward physical model or proxy model, and outputs the predicted surface geometric model.

[0075] The difference calculation unit obtains the surface geometry model at the next actual observation time point (generated by the dynamic surface modeling module) and compares it with the predicted surface geometry model output by the forward prediction unit, calculating the predicted-observed difference between the two (for example, calculating the displacement difference vectors of corresponding vertices). In this embodiment, the predicted-observed difference refers to the difference between the predicted surface geometry model and the actually observed surface geometry model.

[0076] The model self-evolution module receives the prediction-observation difference output by the difference calculation unit of the prediction and difference analysis module.

[0077] In this embodiment, a basic online update strategy is adopted. For example, the difference is used as an error signal, and some model parameters of the subsurface state inversion module (DSIM) are fine-tuned through the gradient descent method (or its variants) to make it better fit the actual observation.

[0078] The causal graph maintenance unit of the causal inference engine module maintains a predefined, simplified causal relationship graph containing the known relationships between basic environmental factors (such as rainfall) and surface deformation.

[0079] The causal effect quantification unit mainly quantifies the impact of factors based on correlation analysis or simple rules (such as the correlation between deformation rate and rainfall intensity).

[0080] The risk attribution analysis unit identifies the main external factors that are temporally related to changes in the current risk status based on the results of the unit.

[0081] The comprehensive assessment unit of the risk assessment and early warning module compares the stress / strain mean field output by the underground state inversion module with the preset rock and soil strength threshold; at the same time, it considers the size of the prediction-observation difference; and combines the basic attribution information output by the causal inference engine module (such as "deformation acceleration is related to rainfall") to evaluate the comprehensive risk level using a method based on thresholds and simple logical rules.

[0082] The warning generation unit generates warning information based on the risk level determined by the comprehensive assessment unit. The information includes the level, approximate location, and a brief explanation of the cause (for example, "accelerated deformation has been monitored, which may be related to recent rainfall").

[0083] Working process

[0084] The data acquisition unit continuously collects the four-dimensional space-time coordinates and optional environmental data of each monitoring point. The data preprocessing unit processes the raw data and outputs a standardized four-dimensional space-time coordinate data set at a fixed time interval. The dynamic surface modeling module updates the surface triangular mesh model M(T) based on the latest coordinate data set. k ), the subsurface state inversion module (DSIM) inputs the time series of the surface model and generates the probabilistic subsurface model S(T k ) (including the mean field and variance field).

[0085] The forward prediction unit uses S(T k ) predicts the surface model M at the next moment pred (T k +ΔT), at T k +ΔT time, the dynamic surface modeling module generates the actual observed surface model M obs (T k +ΔT), the difference calculation unit calculates the difference D(T k +ΔT).

[0086] The model self-evolution module uses the difference D(T k +ΔT) to fine-tune the parameters of the subsurface state inversion module (DSIM), and the causal inference engine module performs basic correlation analysis and attribution on observation data, inversion results, differences and environmental data.

[0087] The comprehensive assessment unit assesses the risk level by combining the inverted underground state (mean), model prediction differences, uncertainty (variance) and basic attribution information. The early warning generation unit issues early warning information containing a brief explanation of the cause based on the risk level. The system repeats this execution at a set time interval (for example, 30 minutes) to achieve dynamic monitoring and early warning.

[0088] It realizes dynamic monitoring and preliminary understanding of underground conditions, overcomes the limitation of traditional monitoring that is limited to surface or point information, and can infer the distribution and uncertainty of key underground mechanical parameters and stress states in real time based on surface dynamic deformation, providing preliminary understanding of the internal state of the mine.

[0089] A basic adaptive capability has been established. By introducing prediction-observation difference feedback and model self-evolution mechanism, the system has a certain self-correction ability and can gradually adapt to the geomechanical characteristics of a specific mine site. Compared with fixed parameter models, it has better long-term applicability.

[0090] The initial introduction of causal thinking, even basic causal analysis, has made the warning information no longer just "exceeding the standard deformation", but can be preliminarily linked to possible external factors (such as rainfall), providing managers with further information to understand the source of risk and uncertainty information: the output probabilistic model (mean and variance field) can quantify the uncertainty of the inversion results, helping decision makers understand the reliability of the current assessment results.

[0091] The construction of a complete system framework has laid the foundation for the subsequent introduction of more AI models, more complex causal inference and more refined risk assessment.

[0092] This embodiment serves as a basic implementation, verifies the feasibility of the entire technical solution, and can provide richer and more dynamic mine safety risk information compared to traditional methods.

[0093] Example 2:

[0094] like Figures 1 to 3 As shown, based on the core system architecture of Example 1, this embodiment focuses on the underground state inversion module and the model self-evolution module, aiming to significantly improve the accuracy and physical reality of underground state inversion and the model's adaptability to dynamic changes in the mine.

[0095] The system hardware deployment and basic data processing flow (data processing module, dynamic surface modeling module) of this embodiment are basically the same as those of the first embodiment. The core difference lies in the implementation of the following modules:

[0096] In this embodiment, the deep learning model built into DSIM specifically uses the physical information-integrated spatiotemporal graph variational autoencoder (PI-STGVAE) architecture, which is precisely designed as follows:

[0097] The spatiotemporal graph encoder (ST-GNNEncoder) receives the time series of surface geometric models output by the dynamic surface modeling module. It uses the graph convolutional network layer to process the spatial structure information of the surface model at each time step, and combines it with recurrent neural network layers (such as GRU) or temporal convolutional layers to capture the evolution of surface deformation over time. It finally outputs a low-dimensional context encoding vector that can represent the dynamic process of the surface.

[0098] Physical Information Integrated Conditional Variational Autoencoder (PI-CVAE);

[0099] Recognition network (Encoder): Input the context encoding vector output by the spatiotemporal graph encoder and output the posterior probability distribution parameters (mean and variance) of the latent variable.

[0100] Latent Space: A low-dimensional space whose vectors represent the essential characteristics of possible subsurface geomechanical states.

[0101] Generator (Decoder): Sampling from the latent space and conditioned on the surface dynamic context encoding vector, it generates a 3D subsurface geomechanical model. This decoder uses structures such as 3D convolutional transpose networks or implicit neural representations (INR) to output probabilistic 3D parameter fields, stress / strain fields, and geological structure representations.

[0102] Physical constraint embedding: During the training and / or inference process of PI-CVAE, its loss function explicitly includes a physical residual term. This term is achieved by randomly selecting configuration points within the generated 3D underground model, using automatic differentiation technology to calculate the residuals of the geomechanical governing equations (including at least the mechanical equilibrium equations or the constitutive relations of the rock and soil) at these points, and minimizing them. This forces the results generated by the model to not only fit the observed data but also to satisfy basic physical laws, thus ensuring the physical authenticity of the inversion results.

[0103] Model training: DSIM is first pre-trained on a large number of synthetic geomechanical simulation datasets covering different geological conditions and mining operations to learn the basic physical laws and inversion mapping relationships. It is then fine-tuned using historical mine monitoring data and limited geological survey information to adapt it to the characteristics of the specific site.

[0104] This embodiment uses an online learning algorithm to update the DSIM. Specifically, at least one of the following can be selected:

[0105] Gradient-based online update: If the prediction process (including forward simulation) is differentiable or approximately differentiable, the gradient of the loss function with respect to the DSIM parameters is calculated using the prediction-observation difference, and the model weights are updated online using stochastic gradient descent (SGD) with a small learning rate or its variants (such as Adam).

[0106] Reinforcement Learning: DSIM is considered an agent whose "action" is to output the underground model S(T_k), the "environment" is the actual evolution process of the mine, the "state" can include historical observations and current inversion results, and the "reward" is based on the size of the prediction-observation difference (the smaller the difference, the higher the reward). Through reinforcement learning algorithms (such as variants of PPO and DQN), the parameters of DSIM are adjusted to make it tend to make inferences that can obtain high rewards (i.e., low prediction errors).

[0107] Through online learning, DSIM can continuously and automatically adapt to the slow or sudden changes in geomechanical behavior caused by mining activities, environmental changes and other factors in the mine.

[0108] The online learning algorithm of the model self-evolution module uses the following formula for adaptive learning rate adjustment factor based on prediction credibility:

[0109]

[0110] Actual update rules:

[0111]

[0112] illustrate:

[0113] : Adaptive learning rate adjustment factor;

[0114] : Base learning rate.

[0115] The difference vector between predictions and observations.

[0116] : Current inversion model The average uncertainty (variance) of .

[0117] and : Normalization factor.

[0118] : Adjustment factor.

[0119] : Hyperbolic tangent activation function, used to limit the adjustment range.

[0120] : are the model parameters at time step t.

[0121] : is the gradient of the loss function with respect to the parameters.

[0122] Working process

[0123] Data collection, preprocessing (data processing module) and dynamic surface modeling (dynamic surface modeling module) are performed in the manner of the first embodiment, and a dynamic surface geometric model sequence is output.

[0124] The PI-STGVAE model of the Subsurface State Inversion Module (DSIM) receives a sequence of surface models and, through its sophisticated encoder-decoder structure and physical constraint mechanism, performs inversion inference to output a more accurate and physically consistent probabilistic three-dimensional subsurface geomechanical model S(T_k).

[0125] The forecast and variance analysis module is based on this high-quality Perform forward-looking predictions to obtain a predicted surface model , and calculate and observe The difference .

[0126] Model self-evolution module receives differences , and applies a selected online learning algorithm (such as gradient-based updates or reinforcement learning) to tune the internal parameters of DSIM.

[0127] The causal inference engine module and the risk assessment and warning module continue to work, but at this time they receive input from DSIM (underground state ) and its uncertainty) of higher quality, model bias information from the prediction and variance analysis modules It can also better reflect the actual model fitting situation.

[0128] The system continues to run in a time-step cycle, and DSIM is continuously optimized under the drive of the self-evolution module.

[0129] The PI-STGVAE architecture can provide a deeper understanding of complex surface spatiotemporal deformation patterns, and combined with physical constraints, it eliminates unreasonable solutions, making the inverted underground structure, parameters, and stress field closer to the actual situation and more reliable.

[0130] By embedding physical laws as hard or soft constraints, the output results are ensured to conform to basic mechanical principles, avoiding the physical fallacies that may arise from purely data-driven models, and enhancing the credibility and engineering practical value of the model.

[0131] The use of a more specific and powerful online learning algorithm for self-evolution enables the system to respond more quickly and effectively to changes in mine conditions, maintain the accuracy of long-term monitoring and early warning, and achieve adaptability far beyond basic implementation.

[0132] Higher-quality knowledge of underground conditions and more reliable model prediction capabilities provide a solid foundation for risk assessment, enabling earlier and more accurate identification of potential risks and reducing false alarm and missed alarm rates.

[0133] The integration of physical information helps the model make relatively reasonable inferences when faced with insufficient monitoring data or encountering new working conditions that are not fully covered in the training data, thereby improving the model's generalization ability.

[0134] This embodiment represents the in-depth optimization of the present invention at the artificial intelligence application level, significantly improving the intelligence level and core performance of the system.

[0135] Example 3:

[0136] like Figures 1 to 3 As shown, based on the first and second embodiments, this embodiment further strengthens the intelligent analysis and decision-making support capabilities of the system, especially in the in-depth tracing of risk causes, intelligent fusion evaluation of multi-dimensional information, and value enhancement of early warning information. In addition, this embodiment also expands the scope of data fusion.

[0137] The system deployment and basic modules of this embodiment (the basic functions of the data processing module, the dynamic surface modeling module, the underground state inversion module using PI-STGVAE, the prediction and difference analysis module, and the model self-evolution module using online learning) are inherited from the first and second embodiments. The core enhancements of this embodiment are as follows:

[0138] In addition to collecting GNSS four-dimensional spatiotemporal coordinate data, the data acquisition unit in this embodiment also integrates data acquisition of at least one surface monitoring technology. For example, it regularly (e.g., every few days) performs drone LiDAR scanning to obtain high-precision three-dimensional point clouds, or connects to a ground-based InSAR system to obtain continuous line-of-sight surface deformation field data.

[0139] The data preprocessing unit adds processing steps such as registration, denoising, and differential calculation (comparison with historical data) to the collected planar monitoring data (such as point clouds and InSAR interferograms), and accurately aligns them with the four-dimensional space-time point data in time and space.

[0140] The dynamic surface modeling module uses the high-precision DSM obtained by drones as the basis for constructing or updating the surface geometric model, improving the initial accuracy and detail expression ability of the model.

[0141] During the training and inference process of its PI-STGVAE model, the Subsurface State Inversion Module (DSIM) incorporates processed surface deformation field data (such as InSAR deformation rate field) as additional observation constraints into the loss function or conditional input, further constraining the inversion results and improving the accuracy of the subsurface model.

[0142] The causal graph maintenance unit of the advanced causal inference engine module no longer relies on a completely predefined causal graph. Instead, it adopts a constraint- or scoring-based causal discovery algorithm (such as the PC algorithm, GES algorithm, etc.), combined with geological background knowledge (as a priori constraints), to learn or dynamically update the causal relationship graph between variables from long-term accumulated multi-source time series data (observations, inversion results, differences, environment, activities).

[0143] The Causal Effect Quantification Unit adopts a more rigorous causal inference method. For example, for controllable or observable interventions (such as blasting and specific excavation steps), a Do-calculus-based method is used to calculate their net causal effect on subsequent stress changes or deformation rates.

[0144] For environmental factors (such as rainfall) or internal changes that are difficult to intervene, a counterfactual model (Potential Outcome Framework) or its approximate method is used to estimate the counterfactual results of "what would the risk indicator be like if the factor did not occur or was at a different level", thereby quantifying its causal contribution.

[0145] Based on the updated causal diagram and quantified causal effects, the risk attribution analysis unit conducts multi-factor, quantitative contribution attribution analysis on monitored abnormal phenomena or assessed high-risk conditions to clarify the primary and secondary relationships of various driving factors.

[0146] The causal effect quantification unit of the causal inference engine module uses the following time-varying causal effect strength formula:

[0147]

[0148] illustrate:

[0149] : time-varying causal effect strength;

[0150] : The maximum instantaneous causal effect of event B on risk indicator Y (such as deformation rate) after the occurrence of event B, estimated by causal inference method;

[0151] : Time when event B occurs;

[0152] : Decay function, for example or other functional forms that take into account the incubation period;

[0153] : The decay parameter or impact time parameter of the causal relationship.

[0154] The comprehensive assessment unit of the intelligent risk assessment and early warning module adopts a more intelligent fusion model, such as a trained meta-learning model (such as a small neural network, gradient boosting decision tree, etc.) or a carefully designed fuzzy logic reasoning system. The model receives the state mean field and variance field (uncertainty) from DSIM, the predicted risk and model bias (difference D) from the prediction and difference analysis module, and most importantly - the various risk drivers and their quantified causal contributions output by the causal inference engine module. The model performs nonlinear weighted fusion on these multi-dimensional information, in which the weight of causal factors is significantly increased, and finally outputs a refined comprehensive risk score or level.

[0155] The warning information generated by the warning generation unit not only includes the risk level, affected area and time elements, but also must include a detailed, quantitative analysis-based causal attribution text description, such as: "Red alert (probability of instability in the next 12 hours > 60%), affecting area XXX.

[0156] The delayed effect of heavy rainfall (accumulated XX mm over the past 48 hours) has led to a significant increase in deep pore water pressure (causal contribution of approximately 55%); the underground inversion model shows that the shear stress on the YYY structural surface has reached a supercritical state (current state risk); recent excavation activities in the ZZZ area have accelerated stress concentration (causal contribution of approximately 20%); 4. The discrepancy between model predictions and observations is widening (decreasing model reliability).

[0157] The comprehensive assessment unit of the risk assessment and early warning module uses the following basic comprehensive risk scoring formula:

[0158]

[0159] illustrate:

[0160] : Basic comprehensive risk score;

[0161] : indicator function, takes value 1 when the condition is met, otherwise 0;

[0162] : mean stress field obtained by inversion;

[0163] : stress threshold;

[0164] : difference vector between prediction and observation;

[0165] : difference threshold;

[0166] : Indicates whether there is a known risk factor (such as heavy rainfall), with a value of 0 or 1;

[0167] : The weight coefficient of each risk signal.

[0168] The comprehensive assessment unit of the risk assessment and early warning module adopts the following causal enhanced risk fusion index formula:

[0169]

[0170] illustrate:

[0171] : Causal Enhanced Risk Fusion Index;

[0172] : A basic risk score calculated based on non-causal factors such as status, uncertainty, and forecast differences;

[0173] : The set of key causal drivers identified;

[0174] : Factor i at time step Normalized causal contribution under (from risk attribution analysis unit);

[0175] : the inherent risk weight of factor i;

[0176] : Causal enhancement index, used to amplify the nonlinear effect of causal factors on risk scores.

[0177] Working process

[0178] The data collection unit collects multi-source data including GNSS four-dimensional spatiotemporal data and at least one type of surface monitoring data.

[0179] The data preprocessing unit processes, aligns and standardizes all data.

[0180] The dynamic surface modeling module uses the fused data to construct a more accurate dynamic surface model M(T_k).

[0181] The Subsurface State Inversion Module (DSIM) uses the PI-STGVAE architecture to perform inversion under the constraints of fused surface data and outputs a more accurate probabilistic subsurface model. .

[0182] The forecast and variance analysis module performs forward forecasts and calculates variances .

[0183] Model self-evolution module utilizes differences Perform online parameter updates on DSIM.

[0184] The causal inference engine modules run in parallel: the unit updates the causal graph, the unit quantifies the causal effects of key factors (such as rainfall, excavation, and internal state changes) using intervention / counterfactual methods, and the unit performs quantitative risk attribution.

[0185] The comprehensive assessment unit inputs the DSIM output (status, uncertainty), the prediction module output (predicted risk, difference D), and the causal engine output (quantitative attribution) into its fusion model (such as fuzzy logic or meta-learning model) to calculate the comprehensive risk level.

[0186] The early warning generation unit generates early warning information containing detailed quantitative causal attribution descriptions based on the risk level and publishes it through the platform.

[0187] The system continues to iterate.

[0188] By introducing causal inference technology, the system can identify the real driving factors of risk evolution and their contribution from complex interactions, going beyond the correlation analysis at the phenomenon level and greatly enhancing the depth of understanding of the internal mechanisms of mine risks.

[0189] The comprehensive assessment unit adopts a more intelligent fusion model and significantly weights causal information, making the risk assessment results closer to the essential driving force of the risk and the assessment more accurate and intelligent.

[0190] The early warning information contains quantitative causal attribution, providing managers with unprecedented decision-making insights, enabling them to clearly determine which link should be prioritized (for example, whether to strengthen drainage, adjust mining plans, or recalibrate models), thereby formulating more effective and targeted emergency response or risk control measures.

[0191] By incorporating planar monitoring data, the spatial coverage and accuracy of the surface deformation field are improved, providing richer constraints for the subsurface state inversion, and further enhancing the accuracy and reliability of the DSIM inversion results.

[0192] The integration of high-precision inversion, adaptive evolution, in-depth cause analysis and intelligent assessment and early warning has enabled the entire system to evolve from a monitoring and early warning tool to a powerful, comprehensive mine safety risk management platform with deep insight and intelligent decision-making support capabilities.

[0193] This embodiment represents an advanced implementation of the present invention in terms of intelligent analysis, causal inference, and comprehensive decision support, and can provide the highest level of technical support for safe production in non-coal mines.

[0194] The above are only preferred embodiments of the present invention. It should be understood that the present invention is not limited to the forms disclosed herein and should not be regarded as excluding other embodiments. Instead, it can be used in various other combinations, modifications and environments, and can be modified within the scope of the conception of this article through the above teachings or technology or knowledge in related fields. The changes and modifications made by those skilled in the art do not depart from the spirit and scope of the present invention and should be within the scope of protection of the claims attached to the present invention.

Claims

1. The non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology is characterized by: It includes data processing module, dynamic surface modeling module, underground state inversion module, prediction and difference analysis module, model self-evolution module, causal inference engine module, risk assessment and early warning module; The data processing module includes at least: a data acquisition unit for collecting non-coal mine monitoring data; a data preprocessing unit for preprocessing the monitoring data to obtain a standardized four-dimensional spatiotemporal data set; a dynamic surface modeling module connected to the data preprocessing unit of the data processing module for constructing a dynamic surface geometric model sequence representing the temporal changes of the surface morphology based on the four-dimensional spatiotemporal data set; The subsurface state inversion module is connected to the dynamic surface modeling module. It is used to perform inversion inference from surface dynamics to subsurface state based on a sequence of dynamic surface geometric models using a deep learning large model to obtain a probabilistic three-dimensional subsurface geomechanical model at the current time step. This model includes estimates of subsurface parameters, stresses, structures, and related uncertainties. The prediction and difference analysis module is connected to the underground state inversion module and the dynamic surface modeling module; the prediction and difference analysis module at least includes: a forward prediction unit for predicting a predicted surface geometry model at the end of a preset time interval in the future based on the currently inverted underground model; a difference calculation unit for obtaining the actual observed surface geometry model at the end of the preset time interval in the future and calculating the prediction-observation difference between the predicted surface geometry model and the actual observed surface geometry model; The model self-evolution module uses the prediction-observation difference as an error signal to adjust the model parameters of the subsurface state inversion module through an online learning algorithm, the algorithm including at least one of gradient-based updating or reinforcement learning; The causal inference engine module connects the dynamic surface modeling module, the subsurface state inversion module, the prediction and differential analysis module, and the module for acquiring environmental / activity data; The risk assessment and early warning module connects the underground state inversion module, the prediction and difference analysis module and the causal inference engine module.

2. The non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology according to claim 1 is characterized by: The monitoring data at least includes four-dimensional spatiotemporal coordinate data of surface monitoring points containing three-dimensional spatial coordinates and corresponding timestamps; and optionally collects surface morphology data, environmental factor data and engineering activity data; preprocessing includes cleaning, denoising, coordinate unification and time synchronization processing. The deep learning large model built into the underground state inversion module is a spatiotemporal graph variational autoencoder architecture that integrates physical information. The architecture includes a spatiotemporal graph encoder for processing the time series of dynamic surface geometric models, and a conditional variational autoencoder for generating a probabilistic three-dimensional underground geomechanical model.

3. The non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology according to claim 2 is characterized by: The causal inference engine module at least includes: a causal graph maintenance unit for maintaining or dynamically learning a graph structure representing the causal relationship between variables; a causal effect quantification unit for quantifying the degree of causal influence of risk-related variables; and a risk attribution analysis unit for identifying key causal drivers that lead to changes in risk status or abnormal phenomena. The probabilistic three-dimensional underground geomechanical model output by the underground state inversion module includes at least: mean estimation and uncertainty quantification provided for the underground medium mechanical parameter field, three-dimensional stress field, and three-dimensional strain field, and the quantification is at least expressed through the variance field; and the probabilistic position representation provided for the geological structure.

4. The non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology according to claim 3 is characterized by: The conditional variational autoencoder that integrates physical information embeds physical law constraints related to geomechanics through a loss function or model structure. The constraints are used to penalize inference results that do not conform to physical laws during training or inference. Physical laws include at least mechanical equilibrium equations or constitutive relations. The risk assessment and early warning module at least includes: a comprehensive assessment unit for integrating the inverted underground state and its uncertainty output by the underground state inversion module, the predicted future risk and model prediction reliability indication output by the prediction and difference analysis module, and the key causal driving factors identified by the causal inference engine module to conduct a comprehensive safety risk level assessment; The warning generation unit is used to generate warnings containing risk levels and causal attribution information based on comprehensive security risk level assessment.

5. The non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology according to claim 4 is characterized by: The dynamic surface modeling module uses a dynamic triangulation algorithm or an adaptive grid generation technology to construct a dynamic surface geometric model based on three-dimensional spatial coordinates and four-dimensional spatiotemporal coordinate data of corresponding timestamps; The model self-evolution module connects the difference calculation unit of the prediction and difference analysis module and the underground state inversion module, and is used to update the model parameters of the underground state inversion module online based on the prediction-observation difference.

6. The non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology according to claim 5 is characterized by: The causal inference engine module is further used to: use a constraint-based or scoring-based causal discovery algorithm to learn or update the causal relationship graph in its causal graph maintenance unit; and adopt an intervention calculus or counterfactual model-based method to quantify the causal effect through its causal effect quantification unit, and perform attribution analysis through its risk attribution analysis unit.

7. The non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology according to claim 6 is characterized by: The comprehensive assessment unit of the risk assessment and early warning module adopts preset rules, fuzzy logic or meta-learning models to perform weighted fusion or logical judgment on the state and uncertainty information from the underground state inversion module, the predicted risk and model deviation information from the prediction and difference analysis module, and the key causal driving factor information from the causal inference engine module to determine the final comprehensive risk level.

8. The non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology according to claim 7 is characterized by: The warning information generated by the warning generation unit of the risk assessment and warning module, in addition to the risk level, location and time elements, also clearly includes a causal attribution text description of the most important one or more reasons leading to the current risk status.

9. The non-coal mine safety risk dynamic monitoring and early warning system based on point-surface fusion technology according to claim 8 is characterized by: The data acquisition unit of the data processing module is also used to collect surface state monitoring data obtained through at least one of ground-based InSAR, UAV LiDAR or UAV photogrammetry technology, and the dynamic surface modeling module and / or underground state inversion module is further used to integrate or utilize surface monitoring data as supplementary information or constraints when constructing a surface model or inverting the underground state.

Images